research25
← Back to research

Research perspective / Machine learning

Using machine learning to target rhamnolipid composition.

A closer look at our published work connecting bioprocess conditions with the composition of biosurfactants.

From bioprocess data to predictive insightStudy workflow / 01–06
  1. 01

    Bioprocess inputs

    Strain, substrate, pH, temperature, cultivation time, and C/N ratio.

  2. 02

    Prepare the dataset

    Organize process conditions alongside observed congener composition.

  3. 03

    Train & compare models

    Random Forest, XGBoost, Gradient Boosting, and neural networks.

  4. 04

    Evaluate predictions

    Compare predicted and observed values to assess model performance.

  5. 05

    Interpret the drivers

    Use SHAP to examine the contribution of each input feature.

  6. 06

    Congener-specific insight

    Connect process conditions with predicted rhamnolipid composition.

Conceptual process-flow diagram based on Nasser, Singh & Kaur (2026). It summarizes the research workflow; it does not reproduce measured results. Read the paper ↗

Beyond total yield.

Rhamnolipids are a family of biosurfactants with different molecular structures. This study asks whether machine learning can help predict their composition from production conditions, rather than focusing only on total output.

Connecting conditions to composition.

The researchers compared four models across nine congener groups. Inputs included substrate, pH, temperature, cultivation time, carbon-to-nitrogen ratio, and bacterial strain. SHAP analysis helped interpret which inputs were influential.

A result with a specific scope.

Random Forest performed best, with a test R² of 0.944 for di-rhamnolipids with two short, saturated chains. Strain, cultivation time, and C/N ratio were prominent predictors in the SHAP analysis.

R² = 0.944

Reported test performance for one congener group—not a blanket score for every prediction.

From a prediction to a useful experiment.

Our perspective: a model is most useful when it helps a team decide what to test next. A practical engagement would begin with the decision at hand, examine the available data, and define how a recommendation could be checked experimentally.

That means asking more than whether a model fits a dataset. How similar are the intended operating conditions to the training examples? What uncertainty matters to the decision? What new evidence would justify moving forward?

These questions connect modelling with process development. They also keep the distinction clear between a promising prediction and a result demonstrated in the process being developed.

Publication
Nasser, M., Singh, A., & Kaur, G. (2026). Machine learning-driven congener-specific rhamnolipids production. Bioresource Technology Reports, article 102518.

DOI: 10.1016/j.biteb.2025.102518 ↗

This article is a plain-language research perspective. Consult the published paper for the methods and complete results.

Working with bioprocess data?

Discuss a modelling or process-development question with Research25.

Email us ↗